github-mirrors/vortex
1
0
Fork 0
mirror of https://github.com/vortexgpgpu/vortex.git synced 2025-04-23 21:39:10 -04:00
Code Issues Projects Releases Packages Wiki Activity Actions

enabling delayed tracing

Browse source
This commit is contained in:
Blaise Tine 2021-08-12 20:05:43 -07:00
parent 7961cf7474
commit c2b3aaa7d1
23 changed files with 305 additions and 274 deletions
Download patch file Download diff file Expand all files Collapse all files

12
driver/opae/vlsim/opae_sim.cpp
View file

@ -10,6 +10,10 @@
#define ENABLE_MEM_STALLS
#ifndef TRACE_DELAY
#define TRACE_DELAY 0
#endif
#ifndef MEM_LATENCY
#define MEM_LATENCY 24
#endif
@ -26,7 +30,9 @@
#define VERILATOR_RESET_VALUE 2
#endif
uint64_t timestamp = 0;
uint64_t sim_trace_delay = TRACE_DELAY;
static uint64_t timestamp = 0;
double sc_time_stamp() {
return timestamp;
@ -198,7 +204,9 @@ void opae_sim::step() {
void opae_sim::eval() {
vortex_afu_->eval();
#ifdef VCD_OUTPUT
trace_->dump(timestamp);
if (timestamp >= sim_trace_delay) {
trace_->dump(timestamp);
}
#endif
++timestamp;
}

44
hw/dpi/float_dpi.vh
View file

@ -1,31 +1,31 @@
`ifndef FLOAT_DPI
`define FLOAT_DPI
import "DPI-C" context function void dpi_fadd(input int a, input int b, input bit[2:0] frm, output int result, output bit[4:0] fflags);
import "DPI-C" context function void dpi_fsub(input int a, input int b, input bit[2:0] frm, output int result, output bit[4:0] fflags);
import "DPI-C" context function void dpi_fmul(input int a, input int b, input bit[2:0] frm, output int result, output bit[4:0] fflags);
import "DPI-C" context function void dpi_fmadd(input int a, input int b, input int c, input bit[2:0] frm, output int result, output bit[4:0] fflags);
import "DPI-C" context function void dpi_fmsub(input int a, input int b, input int c, input bit[2:0] frm, output int result, output bit[4:0] fflags);
import "DPI-C" context function void dpi_fnmadd(input int a, input int b, input int c, input bit[2:0] frm, output int result, output bit[4:0] fflags);
import "DPI-C" context function void dpi_fnmsub(input int a, input int b, input int c, input bit[2:0] frm, output int result, output bit[4:0] fflags);
import "DPI-C" function void dpi_fadd(input int a, input int b, input bit[2:0] frm, output int result, output bit[4:0] fflags);
import "DPI-C" function void dpi_fsub(input int a, input int b, input bit[2:0] frm, output int result, output bit[4:0] fflags);
import "DPI-C" function void dpi_fmul(input int a, input int b, input bit[2:0] frm, output int result, output bit[4:0] fflags);
import "DPI-C" function void dpi_fmadd(input int a, input int b, input int c, input bit[2:0] frm, output int result, output bit[4:0] fflags);
import "DPI-C" function void dpi_fmsub(input int a, input int b, input int c, input bit[2:0] frm, output int result, output bit[4:0] fflags);
import "DPI-C" function void dpi_fnmadd(input int a, input int b, input int c, input bit[2:0] frm, output int result, output bit[4:0] fflags);
import "DPI-C" function void dpi_fnmsub(input int a, input int b, input int c, input bit[2:0] frm, output int result, output bit[4:0] fflags);
import "DPI-C" context function void dpi_fdiv(input int a, input int b, input bit[2:0] frm, output int result, output bit[4:0] fflags);
import "DPI-C" context function void dpi_fsqrt(input int a, input bit[2:0] frm, output int result, output bit[4:0] fflags);
import "DPI-C" function void dpi_fdiv(input int a, input int b, input bit[2:0] frm, output int result, output bit[4:0] fflags);
import "DPI-C" function void dpi_fsqrt(input int a, input bit[2:0] frm, output int result, output bit[4:0] fflags);
import "DPI-C" context function void dpi_ftoi(input int a, input bit[2:0] frm, output int result, output bit[4:0] fflags);
import "DPI-C" context function void dpi_ftou(input int a, input bit[2:0] frm, output int result, output bit[4:0] fflags);
import "DPI-C" context function void dpi_itof(input int a, input bit[2:0] frm, output int result, output bit[4:0] fflags);
import "DPI-C" context function void dpi_utof(input int a, input bit[2:0] frm, output int result, output bit[4:0] fflags);
import "DPI-C" function void dpi_ftoi(input int a, input bit[2:0] frm, output int result, output bit[4:0] fflags);
import "DPI-C" function void dpi_ftou(input int a, input bit[2:0] frm, output int result, output bit[4:0] fflags);
import "DPI-C" function void dpi_itof(input int a, input bit[2:0] frm, output int result, output bit[4:0] fflags);
import "DPI-C" function void dpi_utof(input int a, input bit[2:0] frm, output int result, output bit[4:0] fflags);
import "DPI-C" context function void dpi_fclss(input int a, output int result);
import "DPI-C" context function void dpi_fsgnj(input int a, input int b, output int result);
import "DPI-C" context function void dpi_fsgnjn(input int a, input int b, output int result);
import "DPI-C" context function void dpi_fsgnjx(input int a, input int b, output int result);
import "DPI-C" function void dpi_fclss(input int a, output int result);
import "DPI-C" function void dpi_fsgnj(input int a, input int b, output int result);
import "DPI-C" function void dpi_fsgnjn(input int a, input int b, output int result);
import "DPI-C" function void dpi_fsgnjx(input int a, input int b, output int result);
import "DPI-C" context function void dpi_flt(input int a, input int b, output int result, output bit[4:0] fflags);
import "DPI-C" context function void dpi_fle(input int a, input int b, output int result, output bit[4:0] fflags);
import "DPI-C" context function void dpi_feq(input int a, input int b, output int result, output bit[4:0] fflags);
import "DPI-C" context function void dpi_fmin(input int a, input int b, output int result, output bit[4:0] fflags);
import "DPI-C" context function void dpi_fmax(input int a, input int b, output int result, output bit[4:0] fflags);
import "DPI-C" function void dpi_flt(input int a, input int b, output int result, output bit[4:0] fflags);
import "DPI-C" function void dpi_fle(input int a, input int b, output int result, output bit[4:0] fflags);
import "DPI-C" function void dpi_feq(input int a, input int b, output int result, output bit[4:0] fflags);
import "DPI-C" function void dpi_fmin(input int a, input int b, output int result, output bit[4:0] fflags);
import "DPI-C" function void dpi_fmax(input int a, input int b, output int result, output bit[4:0] fflags);
`endif

16
hw/dpi/util_dpi.cpp
View file

@ -14,8 +14,14 @@ extern "C" {
int dpi_register();
void dpi_assert(int inst, bool cond, int delay);
void dpi_trace(const char* format, ...);
}
double sc_time_stamp();
extern uint64_t sim_trace_delay;
class ShiftRegister {
public:
ShiftRegister() : init_(false), depth_(0) {}
@ -133,4 +139,14 @@ void dpi_idiv(int a, int b, bool is_signed, int* quotient, int* remainder) {
*remainder = dividen % divisor;
}
}
}
void dpi_trace(const char* format, ...) {
uint64_t timestamp = (uint64_t)sc_time_stamp();
if (timestamp < sim_trace_delay)
return;
va_list va;
va_start(va, format);
vprintf(format, va);
va_end(va);
}

10
hw/dpi/util_dpi.vh
View file

@ -1,10 +1,12 @@
`ifndef UTIL_DPI
`define UTIL_DPI
import "DPI-C" context function void dpi_imul(input int a, input int b, input logic is_signed_a, input logic is_signed_b, output int resultl, output int resulth);
import "DPI-C" context function void dpi_idiv(input int a, input int b, input logic is_signed, output int quotient, output int remainder);
import "DPI-C" function void dpi_imul(input int a, input int b, input logic is_signed_a, input logic is_signed_b, output int resultl, output int resulth);
import "DPI-C" function void dpi_idiv(input int a, input int b, input logic is_signed, output int quotient, output int remainder);
import "DPI-C" context function int dpi_register();
import "DPI-C" context function void dpi_assert(int inst, input logic cond, input int delay);
import "DPI-C" function int dpi_register();
import "DPI-C" function void dpi_assert(int inst, input logic cond, input int delay);
import "DPI-C" function void dpi_trace(input string format /*verilator sformat*/);
`endif

4
hw/rtl/VX_alu_unit.v
View file

@ -204,8 +204,8 @@ module VX_alu_unit #(
`ifdef DBG_PRINT_PIPELINE
always @(posedge clk) begin
if (branch_ctl_if.valid) begin
$display("%t: core%0d-branch: wid=%0d, PC=%0h, taken=%b, dest=%0h", $time, CORE_ID,
branch_ctl_if.wid, alu_commit_if.PC, branch_ctl_if.taken, branch_ctl_if.dest);
dpi_trace("%d: core%0d-branch: wid=%0d, PC=%0h, taken=%b, dest=%0h\n",
$time, CORE_ID, branch_ctl_if.wid, alu_commit_if.PC, branch_ctl_if.taken, branch_ctl_if.dest);
end
end
`endif

32
hw/rtl/VX_commit.v
View file

@ -93,34 +93,34 @@ module VX_commit #(
`ifdef DBG_PRINT_PIPELINE
always @(posedge clk) begin
if (alu_commit_if.valid && alu_commit_if.ready) begin
$write("%t: core%0d-commit: wid=%0d, PC=%0h, ex=ALU, tmask=%b, wb=%0d, rd=%0d, data=", $time, CORE_ID, alu_commit_if.wid, alu_commit_if.PC, alu_commit_if.tmask, alu_commit_if.wb, alu_commit_if.rd);
`PRINT_ARRAY1D(alu_commit_if.data, `NUM_THREADS);
$write("\n");
dpi_trace("%d: core%0d-commit: wid=%0d, PC=%0h, ex=ALU, tmask=%b, wb=%0d, rd=%0d, data=", $time, CORE_ID, alu_commit_if.wid, alu_commit_if.PC, alu_commit_if.tmask, alu_commit_if.wb, alu_commit_if.rd);
`TRACE_ARRAY1D(alu_commit_if.data, `NUM_THREADS);
dpi_trace("\n");
end
if (ld_commit_if.valid && ld_commit_if.ready) begin
$write("%t: core%0d-commit: wid=%0d, PC=%0h, ex=LSU, tmask=%b, wb=%0d, rd=%0d, data=", $time, CORE_ID, ld_commit_if.wid, ld_commit_if.PC, ld_commit_if.tmask, ld_commit_if.wb, ld_commit_if.rd);
`PRINT_ARRAY1D(ld_commit_if.data, `NUM_THREADS);
$write("\n");
dpi_trace("%d: core%0d-commit: wid=%0d, PC=%0h, ex=LSU, tmask=%b, wb=%0d, rd=%0d, data=", $time, CORE_ID, ld_commit_if.wid, ld_commit_if.PC, ld_commit_if.tmask, ld_commit_if.wb, ld_commit_if.rd);
`TRACE_ARRAY1D(ld_commit_if.data, `NUM_THREADS);
dpi_trace("\n");
end
if (st_commit_if.valid && st_commit_if.ready) begin
$display("%t: core%0d-commit: wid=%0d, PC=%0h, ex=LSU, tmask=%b, wb=%0d, rd=%0d", $time, CORE_ID, st_commit_if.wid, st_commit_if.PC, st_commit_if.tmask, st_commit_if.wb, st_commit_if.rd);
dpi_trace("%d: core%0d-commit: wid=%0d, PC=%0h, ex=LSU, tmask=%b, wb=%0d, rd=%0d\n", $time, CORE_ID, st_commit_if.wid, st_commit_if.PC, st_commit_if.tmask, st_commit_if.wb, st_commit_if.rd);
end
if (csr_commit_if.valid && csr_commit_if.ready) begin
$write("%t: core%0d-commit: wid=%0d, PC=%0h, ex=CSR, tmask=%b, wb=%0d, rd=%0d, data=", $time, CORE_ID, csr_commit_if.wid, csr_commit_if.PC, csr_commit_if.tmask, csr_commit_if.wb, csr_commit_if.rd);
`PRINT_ARRAY1D(csr_commit_if.data, `NUM_THREADS);
$write("\n");
dpi_trace("%d: core%0d-commit: wid=%0d, PC=%0h, ex=CSR, tmask=%b, wb=%0d, rd=%0d, data=", $time, CORE_ID, csr_commit_if.wid, csr_commit_if.PC, csr_commit_if.tmask, csr_commit_if.wb, csr_commit_if.rd);
`TRACE_ARRAY1D(csr_commit_if.data, `NUM_THREADS);
dpi_trace("\n");
end
`ifdef EXT_F_ENABLE
if (fpu_commit_if.valid && fpu_commit_if.ready) begin
$write("%t: core%0d-commit: wid=%0d, PC=%0h, ex=FPU, tmask=%b, wb=%0d, rd=%0d, data=", $time, CORE_ID, fpu_commit_if.wid, fpu_commit_if.PC, fpu_commit_if.tmask, fpu_commit_if.wb, fpu_commit_if.rd);
`PRINT_ARRAY1D(fpu_commit_if.data, `NUM_THREADS);
$write("\n");
dpi_trace("%d: core%0d-commit: wid=%0d, PC=%0h, ex=FPU, tmask=%b, wb=%0d, rd=%0d, data=", $time, CORE_ID, fpu_commit_if.wid, fpu_commit_if.PC, fpu_commit_if.tmask, fpu_commit_if.wb, fpu_commit_if.rd);
`TRACE_ARRAY1D(fpu_commit_if.data, `NUM_THREADS);
dpi_trace("\n");
end
`endif
if (gpu_commit_if.valid && gpu_commit_if.ready) begin
$write("%t: core%0d-commit: wid=%0d, PC=%0h, ex=GPU, tmask=%b, wb=%0d, rd=%0d, data=", $time, CORE_ID, gpu_commit_if.wid, gpu_commit_if.PC, gpu_commit_if.tmask, gpu_commit_if.wb, gpu_commit_if.rd);
`PRINT_ARRAY1D(gpu_commit_if.data, `NUM_THREADS);
$write("\n");
dpi_trace("%d: core%0d-commit: wid=%0d, PC=%0h, ex=GPU, tmask=%b, wb=%0d, rd=%0d, data=", $time, CORE_ID, gpu_commit_if.wid, gpu_commit_if.PC, gpu_commit_if.tmask, gpu_commit_if.wb, gpu_commit_if.rd);
`TRACE_ARRAY1D(gpu_commit_if.data, `NUM_THREADS);
dpi_trace("\n");
end
end
`endif

6
hw/rtl/VX_decode.v
View file

@ -416,11 +416,11 @@ module VX_decode #(
`ifdef DBG_PRINT_PIPELINE
always @(posedge clk) begin
if (decode_if.valid && decode_if.ready) begin
$write("%t: core%0d-decode: wid=%0d, PC=%0h, ex=", $time, CORE_ID, decode_if.wid, decode_if.PC);
dpi_trace("%d: core%0d-decode: wid=%0d, PC=%0h, ex=", $time, CORE_ID, decode_if.wid, decode_if.PC);
print_ex_type(decode_if.ex_type);
$write(", op=");
dpi_trace(", op=");
print_ex_op(decode_if.ex_type, decode_if.op_type, decode_if.op_mod);
$write(", mod=%0d, tmask=%b, wb=%b, rd=%0d, rs1=%0d, rs2=%0d, rs3=%0d, imm=%0h, use_pc=%b, use_imm=%b, use_regs=%b\n", decode_if.op_mod, decode_if.tmask, decode_if.wb, decode_if.rd, decode_if.rs1, decode_if.rs2, decode_if.rs3, decode_if.imm, decode_if.use_PC, decode_if.use_imm, decode_if.used_regs);
dpi_trace(", mod=%0d, tmask=%b, wb=%b, rd=%0d, rs1=%0d, rs2=%0d, rs3=%0d, imm=%0h, use_pc=%b, use_imm=%b, use_regs=%b\n", decode_if.op_mod, decode_if.tmask, decode_if.wb, decode_if.rd, decode_if.rs1, decode_if.rs2, decode_if.rs3, decode_if.imm, decode_if.use_PC, decode_if.use_imm, decode_if.used_regs);
end
end
`endif

4
hw/rtl/VX_icache_stage.v
View file

@ -92,10 +92,10 @@ module VX_icache_stage #(
`ifdef DBG_PRINT_CORE_ICACHE
always @(posedge clk) begin
if (icache_req_if.valid && icache_req_if.ready) begin
$display("%t: I$%0d req: wid=%0d, PC=%0h", $time, CORE_ID, ifetch_req_if.wid, ifetch_req_if.PC);
dpi_trace("%d: I$%0d req: wid=%0d, PC=%0h\n", $time, CORE_ID, ifetch_req_if.wid, ifetch_req_if.PC);
end
if (ifetch_rsp_if.valid && ifetch_rsp_if.ready) begin
$display("%t: I$%0d rsp: wid=%0d, PC=%0h, data=%0h", $time, CORE_ID, ifetch_rsp_if.wid, ifetch_rsp_if.PC, ifetch_rsp_if.data);
dpi_trace("%d: I$%0d rsp: wid=%0d, PC=%0h, data=%0h\n", $time, CORE_ID, ifetch_rsp_if.wid, ifetch_rsp_if.PC, ifetch_rsp_if.data);
end
end
`endif

50
hw/rtl/VX_issue.v
View file

@ -186,46 +186,46 @@ module VX_issue #(
`ifdef DBG_PRINT_PIPELINE
always @(posedge clk) begin
if (alu_req_if.valid && alu_req_if.ready) begin
$write("%t: core%0d-issue: wid=%0d, PC=%0h, ex=ALU, tmask=%b, rd=%0d, rs1_data=",
dpi_trace("%d: core%0d-issue: wid=%0d, PC=%0h, ex=ALU, tmask=%b, rd=%0d, rs1_data=",
$time, CORE_ID, alu_req_if.wid, alu_req_if.PC, alu_req_if.tmask, alu_req_if.rd);
`PRINT_ARRAY1D(alu_req_if.rs1_data, `NUM_THREADS);
$write(", rs2_data=");
`PRINT_ARRAY1D(alu_req_if.rs2_data, `NUM_THREADS);
$write("\n");
`TRACE_ARRAY1D(alu_req_if.rs1_data, `NUM_THREADS);
dpi_trace(", rs2_data=");
`TRACE_ARRAY1D(alu_req_if.rs2_data, `NUM_THREADS);
dpi_trace("\n");
end
if (lsu_req_if.valid && lsu_req_if.ready) begin
$write("%t: core%0d-issue: wid=%0d, PC=%0h, ex=LSU, tmask=%b, rd=%0d, offset=%0h, addr=",
dpi_trace("%d: core%0d-issue: wid=%0d, PC=%0h, ex=LSU, tmask=%b, rd=%0d, offset=%0h, addr=",
$time, CORE_ID, lsu_req_if.wid, lsu_req_if.PC, lsu_req_if.tmask, lsu_req_if.rd, lsu_req_if.offset);
`PRINT_ARRAY1D(lsu_req_if.base_addr, `NUM_THREADS);
$write(", data=");
`PRINT_ARRAY1D(lsu_req_if.store_data, `NUM_THREADS);
$write("\n");
`TRACE_ARRAY1D(lsu_req_if.base_addr, `NUM_THREADS);
dpi_trace(", data=");
`TRACE_ARRAY1D(lsu_req_if.store_data, `NUM_THREADS);
dpi_trace("\n");
end
if (csr_req_if.valid && csr_req_if.ready) begin
$write("%t: core%0d-issue: wid=%0d, PC=%0h, ex=CSR, tmask=%b, rd=%0d, addr=%0h, rs1_data=",
dpi_trace("%d: core%0d-issue: wid=%0d, PC=%0h, ex=CSR, tmask=%b, rd=%0d, addr=%0h, rs1_data=",
$time, CORE_ID, csr_req_if.wid, csr_req_if.PC, csr_req_if.tmask, csr_req_if.rd, csr_req_if.addr);
`PRINT_ARRAY1D(csr_req_if.rs1_data, `NUM_THREADS);
$write("\n");
`TRACE_ARRAY1D(csr_req_if.rs1_data, `NUM_THREADS);
dpi_trace("\n");
end
`ifdef EXT_F_ENABLE
if (fpu_req_if.valid && fpu_req_if.ready) begin
$write("%t: core%0d-issue: wid=%0d, PC=%0h, ex=FPU, tmask=%b, rd=%0d, rs1_data=",
dpi_trace("%d: core%0d-issue: wid=%0d, PC=%0h, ex=FPU, tmask=%b, rd=%0d, rs1_data=",
$time, CORE_ID, fpu_req_if.wid, fpu_req_if.PC, fpu_req_if.tmask, fpu_req_if.rd);
`PRINT_ARRAY1D(fpu_req_if.rs1_data, `NUM_THREADS);
$write(", rs2_data=");
`PRINT_ARRAY1D(fpu_req_if.rs2_data, `NUM_THREADS);
$write(", rs3_data=");
`PRINT_ARRAY1D(fpu_req_if.rs3_data, `NUM_THREADS);
$write("\n");
`TRACE_ARRAY1D(fpu_req_if.rs1_data, `NUM_THREADS);
dpi_trace(", rs2_data=");
`TRACE_ARRAY1D(fpu_req_if.rs2_data, `NUM_THREADS);
dpi_trace(", rs3_data=");
`TRACE_ARRAY1D(fpu_req_if.rs3_data, `NUM_THREADS);
dpi_trace("\n");
end
`endif
if (gpu_req_if.valid && gpu_req_if.ready) begin
$write("%t: core%0d-issue: wid=%0d, PC=%0h, ex=GPU, tmask=%b, rd=%0d, rs1_data=",
dpi_trace("%d: core%0d-issue: wid=%0d, PC=%0h, ex=GPU, tmask=%b, rd=%0d, rs1_data=",
$time, CORE_ID, gpu_req_if.wid, gpu_req_if.PC, gpu_req_if.tmask, gpu_req_if.rd);
`PRINT_ARRAY1D(gpu_req_if.rs1_data, `NUM_THREADS);
$write(", rs2_data=");
`PRINT_ARRAY1D(gpu_req_if.rs2_data, `NUM_THREADS);
$write("\n");
`TRACE_ARRAY1D(gpu_req_if.rs1_data, `NUM_THREADS);
dpi_trace(", rs2_data=");
`TRACE_ARRAY1D(gpu_req_if.rs2_data, `NUM_THREADS);
dpi_trace("\n");
end
end
`endif

38
hw/rtl/VX_lsu_unit.v
View file

@ -324,8 +324,8 @@ module VX_lsu_unit #(
for (integer i = 0; i < `LSUQ_SIZE; ++i) begin
if (pending_reqs[i][0]) begin
assert(($time - pending_reqs[i][1 +: 64]) < delay_timeout) else
$error("%t: *** D$%0d response timeout: remaining=%b, wid=%0d, PC=%0h, rd=%0d", $time, CORE_ID,
rsp_rem_mask[i], pending_reqs[i][1+64+32+`NR_BITS +: `NW_BITS], pending_reqs[i][1+64+`NR_BITS +: 32], pending_reqs[i][1+64 +: `NR_BITS]);
$error("%t: *** D$%0d response timeout: remaining=%b, wid=%0d, PC=%0h, rd=%0d",
$time, CORE_ID, rsp_rem_mask[i], pending_reqs[i][1+64+32+`NR_BITS +: `NW_BITS], pending_reqs[i][1+64+`NR_BITS +: 32], pending_reqs[i][1+64 +: `NR_BITS]);
end
end
end
@ -333,31 +333,31 @@ module VX_lsu_unit #(
`ifdef DBG_PRINT_CORE_DCACHE
always @(posedge clk) begin
if (lsu_req_if.valid && fence_wait) begin
$display("%t: *** D$%0d fence wait", $time, CORE_ID);
if (lsu_req_if.valid && fence_wait) begin
dpi_trace("%d: *** D$%0d fence wait\n", $time, CORE_ID);
end
if (dcache_req_fire_any) begin
if (dcache_req_if.rw[0]) begin
$write("%t: D$%0d Wr Req: wid=%0d, PC=%0h, tmask=%b, addr=", $time, CORE_ID, req_wid, req_pc, dcache_req_fire);
`PRINT_ARRAY1D(req_addr, `NUM_THREADS);
$write(", tag=%0h, byteen=%0h, type=", req_tag, dcache_req_if.byteen);
`PRINT_ARRAY1D(req_addr_type, `NUM_THREADS);
$write(", data=");
`PRINT_ARRAY1D(dcache_req_if.data, `NUM_THREADS);
$write("\n");
dpi_trace("%d: D$%0d Wr Req: wid=%0d, PC=%0h, tmask=%b, addr=", $time, CORE_ID, req_wid, req_pc, dcache_req_fire);
`TRACE_ARRAY1D(req_addr, `NUM_THREADS);
dpi_trace(", tag=%0h, byteen=%0h, type=", req_tag, dcache_req_if.byteen);
`TRACE_ARRAY1D(req_addr_type, `NUM_THREADS);
dpi_trace(", data=");
`TRACE_ARRAY1D(dcache_req_if.data, `NUM_THREADS);
dpi_trace("\n");
end else begin
$write("%t: D$%0d Rd Req: wid=%0d, PC=%0h, tmask=%b, addr=", $time, CORE_ID, req_wid, req_pc, dcache_req_fire);
`PRINT_ARRAY1D(req_addr, `NUM_THREADS);
$write(", tag=%0h, byteen=%0h, type=", req_tag, dcache_req_if.byteen);
`PRINT_ARRAY1D(req_addr_type, `NUM_THREADS);
$write(", rd=%0d, is_dup=%b\n", req_rd, req_is_dup);
dpi_trace("%d: D$%0d Rd Req: wid=%0d, PC=%0h, tmask=%b, addr=", $time, CORE_ID, req_wid, req_pc, dcache_req_fire);
`TRACE_ARRAY1D(req_addr, `NUM_THREADS);
dpi_trace(", tag=%0h, byteen=%0h, type=", req_tag, dcache_req_if.byteen);
`TRACE_ARRAY1D(req_addr_type, `NUM_THREADS);
dpi_trace(", rd=%0d, is_dup=%b\n", req_rd, req_is_dup);
end
end
if (dcache_rsp_fire) begin
$write("%t: D$%0d Rsp: wid=%0d, PC=%0h, tmask=%b, tag=%0h, rd=%0d, data=",
dpi_trace("%d: D$%0d Rsp: wid=%0d, PC=%0h, tmask=%b, tag=%0h, rd=%0d, data=",
$time, CORE_ID, rsp_wid, rsp_pc, dcache_rsp_if.tmask, mbuf_raddr, rsp_rd);
`PRINT_ARRAY1D(dcache_rsp_if.data, `NUM_THREADS);
$write(", is_dup=%b\n", rsp_is_dup);
`TRACE_ARRAY1D(dcache_rsp_if.data, `NUM_THREADS);
dpi_trace(", is_dup=%b\n", rsp_is_dup);
end
end
`endif

4
hw/rtl/VX_muldiv.v
View file

@ -1,9 +1,5 @@
`include "VX_define.vh"
`ifndef SYNTHESIS
`include "util_dpi.vh"
`endif
module VX_muldiv (
input wire clk,
input wire reset,

30
hw/rtl/VX_platform.vh
View file

@ -1,6 +1,10 @@
`ifndef VX_PLATFORM
`define VX_PLATFORM
`ifndef SYNTHESIS
`include "util_dpi.vh"
`endif
`include "VX_scope.vh"
///////////////////////////////////////////////////////////////////////////////
@ -91,25 +95,25 @@
`define LTRIM(x,s) x[s-1:0]
`define PRINT_ARRAY1D(a, m) \
$write("{"); \
`define TRACE_ARRAY1D(a, m) \
dpi_trace("{"); \
for (integer i = (m-1); i >= 0; --i) begin \
if (i != (m-1)) $write(", "); \
$write("0x%0h", a[i]); \
if (i != (m-1)) dpi_trace(", "); \
dpi_trace("0x%0h", a[i]); \
end \
$write("}"); \
dpi_trace("}"); \
`define PRINT_ARRAY2D(a, m, n) \
$write("{"); \
`define TRACE_ARRAY2D(a, m, n) \
dpi_trace("{"); \
for (integer i = n-1; i >= 0; --i) begin \
if (i != (n-1)) $write(", "); \
$write("{"); \
if (i != (n-1)) dpi_trace(", "); \
dpi_trace("{"); \
for (integer j = (m-1); j >= 0; --j) begin \
if (j != (m-1)) $write(", "); \
$write("0x%0h", a[i][j]); \
if (j != (m-1)) dpi_trace(", "); \
dpi_trace("0x%0h", a[i][j]); \
end \
$write("}"); \
dpi_trace("}"); \
end \
$write("}")
dpi_trace("}")
`endif

168
hw/rtl/VX_print_instr.vh
View file

@ -7,12 +7,12 @@ task print_ex_type (
input [`EX_BITS-1:0] ex_type
);
case (ex_type)
`EX_ALU: $write("ALU");
`EX_LSU: $write("LSU");
`EX_CSR: $write("CSR");
`EX_FPU: $write("FPU");
`EX_GPU: $write("GPU");
default: $write("NOP");
`EX_ALU: dpi_trace("ALU");
`EX_LSU: dpi_trace("LSU");
`EX_CSR: dpi_trace("CSR");
`EX_FPU: dpi_trace("FPU");
`EX_GPU: dpi_trace("GPU");
default: dpi_trace("NOP");
endcase
endtask
@ -25,113 +25,113 @@ task print_ex_op (
`EX_ALU: begin
if (`ALU_IS_BR(op_mod)) begin
case (`BR_BITS'(op_type))
`BR_EQ: $write("BEQ");
`BR_NE: $write("BNE");
`BR_LT: $write("BLT");
`BR_GE: $write("BGE");
`BR_LTU: $write("BLTU");
`BR_GEU: $write("BGEU");
`BR_JAL: $write("JAL");
`BR_JALR: $write("JALR");
`BR_ECALL: $write("ECALL");
`BR_EBREAK:$write("EBREAK");
`BR_MRET: $write("MRET");
`BR_SRET: $write("SRET");
`BR_DRET: $write("DRET");
default: $write("?");
`BR_EQ: dpi_trace("BEQ");
`BR_NE: dpi_trace("BNE");
`BR_LT: dpi_trace("BLT");
`BR_GE: dpi_trace("BGE");
`BR_LTU: dpi_trace("BLTU");
`BR_GEU: dpi_trace("BGEU");
`BR_JAL: dpi_trace("JAL");
`BR_JALR: dpi_trace("JALR");
`BR_ECALL: dpi_trace("ECALL");
`BR_EBREAK:dpi_trace("EBREAK");
`BR_MRET: dpi_trace("MRET");
`BR_SRET: dpi_trace("SRET");
`BR_DRET: dpi_trace("DRET");
default: dpi_trace("?");
endcase
end else if (`ALU_IS_MUL(op_mod)) begin
case (`MUL_BITS'(op_type))
`MUL_MUL: $write("MUL");
`MUL_MULH: $write("MULH");
`MUL_MULHSU:$write("MULHSU");
`MUL_MULHU: $write("MULHU");
`MUL_DIV: $write("DIV");
`MUL_DIVU: $write("DIVU");
`MUL_REM: $write("REM");
`MUL_REMU: $write("REMU");
default: $write("?");
`MUL_MUL: dpi_trace("MUL");
`MUL_MULH: dpi_trace("MULH");
`MUL_MULHSU:dpi_trace("MULHSU");
`MUL_MULHU: dpi_trace("MULHU");
`MUL_DIV: dpi_trace("DIV");
`MUL_DIVU: dpi_trace("DIVU");
`MUL_REM: dpi_trace("REM");
`MUL_REMU: dpi_trace("REMU");
default: dpi_trace("?");
endcase
end else begin
case (`ALU_BITS'(op_type))
`ALU_ADD: $write("ADD");
`ALU_SUB: $write("SUB");
`ALU_SLL: $write("SLL");
`ALU_SRL: $write("SRL");
`ALU_SRA: $write("SRA");
`ALU_SLT: $write("SLT");
`ALU_SLTU: $write("SLTU");
`ALU_XOR: $write("XOR");
`ALU_OR: $write("OR");
`ALU_AND: $write("AND");
`ALU_LUI: $write("LUI");
`ALU_AUIPC: $write("AUIPC");
default: $write("?");
`ALU_ADD: dpi_trace("ADD");
`ALU_SUB: dpi_trace("SUB");
`ALU_SLL: dpi_trace("SLL");
`ALU_SRL: dpi_trace("SRL");
`ALU_SRA: dpi_trace("SRA");
`ALU_SLT: dpi_trace("SLT");
`ALU_SLTU: dpi_trace("SLTU");
`ALU_XOR: dpi_trace("XOR");
`ALU_OR: dpi_trace("OR");
`ALU_AND: dpi_trace("AND");
`ALU_LUI: dpi_trace("LUI");
`ALU_AUIPC: dpi_trace("AUIPC");
default: dpi_trace("?");
endcase
end
end
`EX_LSU: begin
case (`LSU_BITS'(op_type))
`LSU_LB: $write("LB");
`LSU_LH: $write("LH");
`LSU_LW: $write("LW");
`LSU_LBU:$write("LBU");
`LSU_LHU:$write("LHU");
`LSU_SB: $write("SB");
`LSU_SH: $write("SH");
`LSU_SW: $write("SW");
default: $write("?");
`LSU_LB: dpi_trace("LB");
`LSU_LH: dpi_trace("LH");
`LSU_LW: dpi_trace("LW");
`LSU_LBU:dpi_trace("LBU");
`LSU_LHU:dpi_trace("LHU");
`LSU_SB: dpi_trace("SB");
`LSU_SH: dpi_trace("SH");
`LSU_SW: dpi_trace("SW");
default: dpi_trace("?");
endcase
end
`EX_CSR: begin
case (`CSR_BITS'(op_type))
`CSR_RW: $write("CSRW");
`CSR_RS: $write("CSRS");
`CSR_RC: $write("CSRC");
default: $write("?");
`CSR_RW: dpi_trace("CSRW");
`CSR_RS: dpi_trace("CSRS");
`CSR_RC: dpi_trace("CSRC");
default: dpi_trace("?");
endcase
end
`EX_FPU: begin
case (`FPU_BITS'(op_type))
`FPU_ADD: $write("ADD");
`FPU_SUB: $write("SUB");
`FPU_MUL: $write("MUL");
`FPU_DIV: $write("DIV");
`FPU_SQRT: $write("SQRT");
`FPU_MADD: $write("MADD");
`FPU_NMSUB: $write("NMSUB");
`FPU_NMADD: $write("NMADD");
`FPU_CVTWS: $write("CVTWS");
`FPU_CVTWUS:$write("CVTWUS");
`FPU_CVTSW: $write("CVTSW");
`FPU_CVTSWU:$write("CVTSWU");
`FPU_CLASS: $write("CLASS");
`FPU_CMP: $write("CMP");
`FPU_ADD: dpi_trace("ADD");
`FPU_SUB: dpi_trace("SUB");
`FPU_MUL: dpi_trace("MUL");
`FPU_DIV: dpi_trace("DIV");
`FPU_SQRT: dpi_trace("SQRT");
`FPU_MADD: dpi_trace("MADD");
`FPU_NMSUB: dpi_trace("NMSUB");
`FPU_NMADD: dpi_trace("NMADD");
`FPU_CVTWS: dpi_trace("CVTWS");
`FPU_CVTWUS:dpi_trace("CVTWUS");
`FPU_CVTSW: dpi_trace("CVTSW");
`FPU_CVTSWU:dpi_trace("CVTSWU");
`FPU_CLASS: dpi_trace("CLASS");
`FPU_CMP: dpi_trace("CMP");
`FPU_MISC: begin
case (op_mod)
0: $write("SGNJ");
1: $write("SGNJN");
2: $write("SGNJX");
3: $write("MIN");
4: $write("MAX");
5: $write("MVXW");
6: $write("MVWX");
0: dpi_trace("SGNJ");
1: dpi_trace("SGNJN");
2: dpi_trace("SGNJX");
3: dpi_trace("MIN");
4: dpi_trace("MAX");
5: dpi_trace("MVXW");
6: dpi_trace("MVWX");
endcase
end
default: $write("?");
default: dpi_trace("?");
endcase
end
`EX_GPU: begin
case (`GPU_BITS'(op_type))
`GPU_TMC: $write("TMC");
`GPU_WSPAWN:$write("WSPAWN");
`GPU_SPLIT: $write("SPLIT");
`GPU_JOIN: $write("JOIN");
`GPU_BAR: $write("BAR");
default: $write("?");
`GPU_TMC: dpi_trace("TMC");
`GPU_WSPAWN:dpi_trace("WSPAWN");
`GPU_SPLIT: dpi_trace("SPLIT");
`GPU_JOIN: dpi_trace("JOIN");
`GPU_BAR: dpi_trace("BAR");
default: dpi_trace("?");
endcase
end
default: $write("?");
default: dpi_trace("?");
endcase
endtask

6
hw/rtl/VX_scoreboard.v
View file

@ -47,9 +47,9 @@ module VX_scoreboard #(
end else begin
`ifdef DBG_PRINT_PIPELINE
if (ibuffer_if.valid && ~ibuffer_if.ready) begin
$display("%t: *** core%0d-stall: wid=%0d, PC=%0h, rd=%0d, wb=%0d, inuse=%b%b%b%b",
$time, CORE_ID, ibuffer_if.wid, ibuffer_if.PC, ibuffer_if.rd, ibuffer_if.wb,
deq_inuse_regs[ibuffer_if.rd], deq_inuse_regs[ibuffer_if.rs1], deq_inuse_regs[ibuffer_if.rs2], deq_inuse_regs[ibuffer_if.rs3]);
dpi_trace("%d: *** core%0d-stall: wid=%0d, PC=%0h, rd=%0d, wb=%0d, inuse=%b%b%b%b\n",
$time, CORE_ID, ibuffer_if.wid, ibuffer_if.PC, ibuffer_if.rd, ibuffer_if.wb,
deq_inuse_regs[ibuffer_if.rd], deq_inuse_regs[ibuffer_if.rs1], deq_inuse_regs[ibuffer_if.rs2], deq_inuse_regs[ibuffer_if.rs3]);
end
`endif
if (release_reg) begin

6
hw/rtl/Vortex.v
View file

@ -201,12 +201,12 @@ module Vortex (
always @(posedge clk) begin
if (mem_req_valid && mem_req_ready) begin
if (mem_req_rw)
$display("%t: MEM Wr Req: addr=%0h, tag=%0h, byteen=%0h data=%0h", $time, `TO_FULL_ADDR(mem_req_addr), mem_req_tag, mem_req_byteen, mem_req_data);
dpi_trace("%d: MEM Wr Req: addr=%0h, tag=%0h, byteen=%0h data=%0h\n", $time, `TO_FULL_ADDR(mem_req_addr), mem_req_tag, mem_req_byteen, mem_req_data);
else
$display("%t: MEM Rd Req: addr=%0h, tag=%0h, byteen=%0h", $time, `TO_FULL_ADDR(mem_req_addr), mem_req_tag, mem_req_byteen);
dpi_trace("%d: MEM Rd Req: addr=%0h, tag=%0h, byteen=%0h\n", $time, `TO_FULL_ADDR(mem_req_addr), mem_req_tag, mem_req_byteen);
end
if (mem_rsp_valid && mem_rsp_ready) begin
$display("%t: MEM Rsp: tag=%0h, data=%0h", $time, mem_rsp_tag, mem_rsp_data);
dpi_trace("%d: MEM Rsp: tag=%0h, data=%0h\n", $time, mem_rsp_tag, mem_rsp_data);
end
end
`endif

11
hw/rtl/afu/VX_avs_wrapper.v
View file

@ -166,13 +166,14 @@ module VX_avs_wrapper #(
`ifdef DBG_PRINT_AVS
always @(posedge clk) begin
if (mem_req_valid && mem_req_ready) begin
if (mem_req_rw)
$display("%t: AVS Wr Req: addr=%0h, byteen=%0h, tag=%0h, data=%0h", $time, `TO_FULL_ADDR(mem_req_addr), mem_req_byteen, mem_req_tag, mem_req_data);
else
$display("%t: AVS Rd Req: addr=%0h, byteen=%0h, tag=%0h, pending=%0d", $time, `TO_FULL_ADDR(mem_req_addr), mem_req_byteen, mem_req_tag, req_queue_size);
if (mem_req_rw) begin
dpi_trace("%d: AVS Wr Req: addr=%0h, byteen=%0h, tag=%0h, data=%0h\n", $time, `TO_FULL_ADDR(mem_req_addr), mem_req_byteen, mem_req_tag, mem_req_data);
end else begin
dpi_trace("%d: AVS Rd Req: addr=%0h, byteen=%0h, tag=%0h, pending=%0d\n", $time, `TO_FULL_ADDR(mem_req_addr), mem_req_byteen, mem_req_tag, req_queue_size);
end
end
if (mem_rsp_valid && mem_rsp_ready) begin
$display("%t: AVS Rd Rsp: tag=%0h, data=%0h, pending=%0d", $time, mem_rsp_tag, mem_rsp_data, req_queue_size);
dpi_trace("%d: AVS Rd Rsp: tag=%0h, data=%0h, pending=%0d\n", $time, mem_rsp_tag, mem_rsp_data, req_queue_size);
end
end
`endif

44
hw/rtl/afu/vortex_afu.sv
View file

@ -186,36 +186,36 @@ always @(posedge clk) begin
MMIO_IO_ADDR: begin
cmd_io_addr <= t_ccip_clAddr'(cp2af_sRxPort.c0.data);
`ifdef DBG_PRINT_OPAE
$display("%t: MMIO_IO_ADDR: addr=%0h, data=0x%0h", $time, mmio_hdr.address, t_ccip_clAddr'(cp2af_sRxPort.c0.data));
dpi_trace("%d: MMIO_IO_ADDR: addr=%0h, data=0x%0h\n", $time, mmio_hdr.address, t_ccip_clAddr'(cp2af_sRxPort.c0.data));
`endif
end
MMIO_MEM_ADDR: begin
cmd_mem_addr <= $bits(cmd_mem_addr)'(cp2af_sRxPort.c0.data);
`ifdef DBG_PRINT_OPAE
$display("%t: MMIO_MEM_ADDR: addr=%0h, data=0x%0h", $time, mmio_hdr.address, $bits(cmd_mem_addr)'(cp2af_sRxPort.c0.data));
dpi_trace("%d: MMIO_MEM_ADDR: addr=%0h, data=0x%0h\n", $time, mmio_hdr.address, $bits(cmd_mem_addr)'(cp2af_sRxPort.c0.data));
`endif
end
MMIO_DATA_SIZE: begin
cmd_data_size <= $bits(cmd_data_size)'(cp2af_sRxPort.c0.data);
`ifdef DBG_PRINT_OPAE
$display("%t: MMIO_DATA_SIZE: addr=%0h, data=%0d", $time, mmio_hdr.address, $bits(cmd_data_size)'(cp2af_sRxPort.c0.data));
dpi_trace("%d: MMIO_DATA_SIZE: addr=%0h, data=%0d\n", $time, mmio_hdr.address, $bits(cmd_data_size)'(cp2af_sRxPort.c0.data));
`endif
end
MMIO_CMD_TYPE: begin
`ifdef DBG_PRINT_OPAE
$display("%t: MMIO_CMD_TYPE: addr=%0h, data=%0d", $time, mmio_hdr.address, $bits(cmd_type)'(cp2af_sRxPort.c0.data));
dpi_trace("%d: MMIO_CMD_TYPE: addr=%0h, data=%0d\n", $time, mmio_hdr.address, $bits(cmd_type)'(cp2af_sRxPort.c0.data));
`endif
end
`ifdef SCOPE
MMIO_SCOPE_WRITE: begin
`ifdef DBG_PRINT_OPAE
$display("%t: MMIO_SCOPE_WRITE: addr=%0h, data=%0h", $time, mmio_hdr.address, 64'(cp2af_sRxPort.c0.data));
dpi_trace("%d: MMIO_SCOPE_WRITE: addr=%0h, data=%0h\n", $time, mmio_hdr.address, 64'(cp2af_sRxPort.c0.data));
`endif
end
`endif
default: begin
`ifdef DBG_PRINT_OPAE
$display("%t: Unknown MMIO Wr: addr=%0h, data=%0h", $time, mmio_hdr.address, $bits(cmd_data_size)'(cp2af_sRxPort.c0.data));
dpi_trace("%d: Unknown MMIO Wr: addr=%0h, data=%0h\n", $time, mmio_hdr.address, $bits(cmd_data_size)'(cp2af_sRxPort.c0.data));
`endif
end
endcase
@ -243,7 +243,7 @@ always @(posedge clk) begin
mmio_tx.data <= 64'({cout_q_dout, !cout_q_empty, 8'(state)});
`ifdef DBG_PRINT_OPAE
if (state != STATE_WIDTH'(mmio_tx.data)) begin
$display("%t: MMIO_STATUS: addr=%0h, state=%0d", $time, mmio_hdr.address, state);
dpi_trace("%d: MMIO_STATUS: addr=%0h, state=%0d\n", $time, mmio_hdr.address, state);
end
`endif
end
@ -251,20 +251,20 @@ always @(posedge clk) begin
MMIO_SCOPE_READ: begin
mmio_tx.data <= cmd_scope_rdata;
`ifdef DBG_PRINT_OPAE
$display("%t: MMIO_SCOPE_READ: addr=%0h, data=%0h", $time, mmio_hdr.address, cmd_scope_rdata);
dpi_trace("%d: MMIO_SCOPE_READ: addr=%0h, data=%0h\n", $time, mmio_hdr.address, cmd_scope_rdata);
`endif
end
`endif
MMIO_DEV_CAPS: begin
mmio_tx.data <= dev_caps;
`ifdef DBG_PRINT_OPAE
$display("%t: MMIO_DEV_CAPS: addr=%0h, data=%0h", $time, mmio_hdr.address, dev_caps);
dpi_trace("%d: MMIO_DEV_CAPS: addr=%0h, data=%0h\n", $time, mmio_hdr.address, dev_caps);
`endif
end
default: begin
mmio_tx.data <= 64'h0;
`ifdef DBG_PRINT_OPAE
$display("%t: Unknown MMIO Rd: addr=%0h", $time, mmio_hdr.address);
dpi_trace("%d: Unknown MMIO Rd: addr=%0h\n", $time, mmio_hdr.address);
`endif
end
endcase
@ -298,19 +298,19 @@ always @(posedge clk) begin
case (cmd_type)
CMD_MEM_READ: begin
`ifdef DBG_PRINT_OPAE
$display("%t: STATE READ: ia=%0h addr=%0h size=%0d", $time, cmd_io_addr, cmd_mem_addr, cmd_data_size);
dpi_trace("%d: STATE READ: ia=%0h addr=%0h size=%0d\n", $time, cmd_io_addr, cmd_mem_addr, cmd_data_size);
`endif
state <= STATE_READ;
end
CMD_MEM_WRITE: begin
`ifdef DBG_PRINT_OPAE
$display("%t: STATE WRITE: ia=%0h addr=%0h size=%0d", $time, cmd_io_addr, cmd_mem_addr, cmd_data_size);
dpi_trace("%d: STATE WRITE: ia=%0h addr=%0h size=%0d\n", $time, cmd_io_addr, cmd_mem_addr, cmd_data_size);
`endif
state <= STATE_WRITE;
end
CMD_RUN: begin
`ifdef DBG_PRINT_OPAE
$display("%t: STATE START", $time);
dpi_trace("%d: STATE START\n", $time);
`endif
vx_reset <= 1;
state <= STATE_START;
@ -325,7 +325,7 @@ always @(posedge clk) begin
if (cmd_read_done) begin
state <= STATE_IDLE;
`ifdef DBG_PRINT_OPAE
$display("%t: STATE IDLE", $time);
dpi_trace("%d: STATE IDLE\n", $time);
`endif
end
end
@ -334,7 +334,7 @@ always @(posedge clk) begin
if (cmd_write_done) begin
state <= STATE_IDLE;
`ifdef DBG_PRINT_OPAE
$display("%t: STATE IDLE", $time);
dpi_trace("%d: STATE IDLE\n", $time);
`endif
end
end
@ -346,7 +346,7 @@ always @(posedge clk) begin
vx_started <= 0;
state <= STATE_IDLE;
`ifdef DBG_PRINT_OPAE
$display("%t: STATE IDLE", $time);
dpi_trace("%d: STATE IDLE\n", $time);
`endif
end
end else begin
@ -697,7 +697,7 @@ always @(posedge clk) begin
cci_rd_req_addr <= cci_rd_req_addr + 1;
cci_rd_req_ctr <= cci_rd_req_ctr + 1;
`ifdef DBG_PRINT_OPAE
$display("%t: CCI Rd Req: addr=%0h, tag=%0h, rem=%0d, pending=%0d", $time, cci_rd_req_addr, cci_rd_req_tag, (cmd_data_size - cci_rd_req_ctr - 1), cci_pending_reads);
dpi_trace("%d: CCI Rd Req: addr=%0h, tag=%0h, rem=%0d, pending=%0d\n", $time, cci_rd_req_addr, cci_rd_req_tag, (cmd_data_size - cci_rd_req_ctr - 1), cci_pending_reads);
`endif
end
@ -707,13 +707,13 @@ always @(posedge clk) begin
cci_mem_wr_req_addr_base <= cci_mem_wr_req_addr_base + CCI_ADDR_WIDTH'(CCI_RD_WINDOW_SIZE);
end
`ifdef DBG_PRINT_OPAE
$display("%t: CCI Rd Rsp: idx=%0d, ctr=%0d, data=%0h", $time, cci_rd_rsp_tag, cci_rd_rsp_ctr, cp2af_sRxPort.c0.data);
dpi_trace("%d: CCI Rd Rsp: idx=%0d, ctr=%0d, data=%0h\n", $time, cci_rd_rsp_tag, cci_rd_rsp_ctr, cp2af_sRxPort.c0.data);
`endif
end
if (cci_rdq_pop) begin
`ifdef DBG_PRINT_OPAE
$display("%t: CCI Rd Queue Pop: pending=%0d", $time, cci_pending_reads);
dpi_trace("%d: CCI Rd Queue Pop: pending=%0d\n", $time, cci_pending_reads);
`endif
end
@ -852,13 +852,13 @@ begin
cci_wr_req_done <= 1;
end
`ifdef DBG_PRINT_OPAE
$display("%t: CCI Wr Req: addr=%0h, rem=%0d, pending=%0d, data=%0h", $time, cci_wr_req_addr, (cci_wr_req_ctr - 1), cci_pending_writes, af2cp_sTxPort.c1.data);
dpi_trace("%d: CCI Wr Req: addr=%0h, rem=%0d, pending=%0d, data=%0h\n", $time, cci_wr_req_addr, (cci_wr_req_ctr - 1), cci_pending_writes, af2cp_sTxPort.c1.data);
`endif
end
if (cci_wr_rsp_fire) begin
`ifdef DBG_PRINT_OPAE
$display("%t: CCI Wr Rsp: pending=%0d", $time, cci_pending_writes);
dpi_trace("%d: CCI Wr Rsp: pending=%0d\n", $time, cci_pending_writes);
`endif
end
end
@ -879,7 +879,7 @@ Vortex #() vortex (
`SCOPE_BIND_afu_vortex
.clk (clk),
.reset (reset | vx_reset),
.reset (vx_reset),
// Memory request
.mem_req_valid (vx_mem_req_valid),

18
hw/rtl/cache/VX_bank.v vendored
View file

@ -527,31 +527,31 @@ module VX_bank #(
always @(posedge clk) begin
if (pipeline_stall) begin
$display("%d: *** cache%0d:%0d stall: crsq=%b, mreq=%b, mshr=%b", $time, CACHE_ID, BANK_ID, crsq_stall, mreq_alm_full, mshr_alm_full);
dpi_trace("%d: *** cache%0d:%0d stall: crsq=%b, mreq=%b, mshr=%b\n", $time, CACHE_ID, BANK_ID, crsq_stall, mreq_alm_full, mshr_alm_full);
end
if (flush_enable) begin
$display("%t: cache%0d:%0d flush: addr=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(flush_addr, BANK_ID));
dpi_trace("%d: cache%0d:%0d flush: addr=%0h\n", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(flush_addr, BANK_ID));
end
if (mem_rsp_fire) begin
$display("%t: cache%0d:%0d fill-rsp: addr=%0h, id=%0d, data=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(mem_rsp_addr, BANK_ID), mem_rsp_id, mem_rsp_data);
dpi_trace("%d: cache%0d:%0d fill-rsp: addr=%0h, id=%0d, data=%0h\n", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(mem_rsp_addr, BANK_ID), mem_rsp_id, mem_rsp_data);
end
if (mshr_fire) begin
$display("%t: cache%0d:%0d mshr-pop: addr=%0h, tag=%0h, pmask=%b, tid=%0d, wid=%0d, PC=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(mshr_addr, BANK_ID), mshr_tag, mshr_pmask, mshr_tid, debug_wid_sel, debug_pc_sel);
dpi_trace("%d: cache%0d:%0d mshr-pop: addr=%0h, tag=%0h, pmask=%b, tid=%0d, wid=%0d, PC=%0h\n", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(mshr_addr, BANK_ID), mshr_tag, mshr_pmask, mshr_tid, debug_wid_sel, debug_pc_sel);
end
if (creq_fire) begin
if (creq_rw)
$display("%t: cache%0d:%0d core-wr-req: addr=%0h, tag=%0h, pmask=%b, tid=%0d, byteen=%b, data=%0h, wid=%0d, PC=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(creq_addr, BANK_ID), creq_tag, creq_pmask, creq_tid, creq_byteen, creq_data, debug_wid_sel, debug_pc_sel);
dpi_trace("%d: cache%0d:%0d core-wr-req: addr=%0h, tag=%0h, pmask=%b, tid=%0d, byteen=%b, data=%0h, wid=%0d, PC=%0h\n", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(creq_addr, BANK_ID), creq_tag, creq_pmask, creq_tid, creq_byteen, creq_data, debug_wid_sel, debug_pc_sel);
else
$display("%t: cache%0d:%0d core-rd-req: addr=%0h, tag=%0h, pmask=%b, tid=%0d, byteen=%b, wid=%0d, PC=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(creq_addr, BANK_ID), creq_tag, creq_pmask, creq_tid, creq_byteen, debug_wid_sel, debug_pc_sel);
dpi_trace("%d: cache%0d:%0d core-rd-req: addr=%0h, tag=%0h, pmask=%b, tid=%0d, byteen=%b, wid=%0d, PC=%0h\n", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(creq_addr, BANK_ID), creq_tag, creq_pmask, creq_tid, creq_byteen, debug_wid_sel, debug_pc_sel);
end
if (crsq_fire) begin
$display("%t: cache%0d:%0d core-rsp: addr=%0h, tag=%0h, pmask=%b, tid=%0d, data=%0h, wid=%0d, PC=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(addr_st1, BANK_ID), crsq_tag, crsq_pmask, crsq_tid, crsq_data, debug_wid_st1, debug_pc_st1);
dpi_trace("%d: cache%0d:%0d core-rsp: addr=%0h, tag=%0h, pmask=%b, tid=%0d, data=%0h, wid=%0d, PC=%0h\n", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(addr_st1, BANK_ID), crsq_tag, crsq_pmask, crsq_tid, crsq_data, debug_wid_st1, debug_pc_st1);
end
if (mreq_push) begin
if (do_writeback_st1)
$display("%t: cache%0d:%0d writeback: addr=%0h, data=%0h, byteen=%b, wid=%0d, PC=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(mreq_addr, BANK_ID), mreq_data, mreq_byteen, debug_wid_st1, debug_pc_st1);
dpi_trace("%d: cache%0d:%0d writeback: addr=%0h, data=%0h, byteen=%b, wid=%0d, PC=%0h\n", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(mreq_addr, BANK_ID), mreq_data, mreq_byteen, debug_wid_st1, debug_pc_st1);
else
$display("%t: cache%0d:%0d fill-req: addr=%0h, id=%0d, wid=%0d, PC=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(mreq_addr, BANK_ID), mreq_id, debug_wid_st1, debug_pc_st1);
dpi_trace("%d: cache%0d:%0d fill-req: addr=%0h, id=%0d, wid=%0d, PC=%0h\n", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(mreq_addr, BANK_ID), mreq_id, debug_wid_st1, debug_pc_st1);
end
end
`endif

6
hw/rtl/cache/VX_data_access.v vendored
View file

@ -83,13 +83,13 @@ module VX_data_access #(
always @(posedge clk) begin
if (writeen && ~stall) begin
if (is_fill) begin
$display("%t: cache%0d:%0d data-fill: addr=%0h, blk_addr=%0d, data=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(addr, BANK_ID), line_addr, wdata);
dpi_trace("%d: cache%0d:%0d data-fill: addr=%0h, blk_addr=%0d, data=%0h\n", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(addr, BANK_ID), line_addr, wdata);
end else begin
$display("%t: cache%0d:%0d data-write: addr=%0h, wid=%0d, PC=%0h, byteen=%b, blk_addr=%0d, data=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(addr, BANK_ID), debug_wid, debug_pc, byte_enable, line_addr, wdata);
dpi_trace("%d: cache%0d:%0d data-write: addr=%0h, wid=%0d, PC=%0h, byteen=%b, blk_addr=%0d, data=%0h\n", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(addr, BANK_ID), debug_wid, debug_pc, byte_enable, line_addr, wdata);
end
end
if (readen && ~stall) begin
$display("%t: cache%0d:%0d data-read: addr=%0h, wid=%0d, PC=%0h, blk_addr=%0d, data=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(addr, BANK_ID), debug_wid, debug_pc, line_addr, rdata);
dpi_trace("%d: cache%0d:%0d data-read: addr=%0h, wid=%0d, PC=%0h, blk_addr=%0d, data=%0h\n", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(addr, BANK_ID), debug_wid, debug_pc, line_addr, rdata);
end
end
`endif

22
hw/rtl/cache/VX_miss_resrv.v vendored
View file

@ -202,33 +202,33 @@ module VX_miss_resrv #(
always @(posedge clk) begin
if (allocate_fire || fill_valid || dequeue_fire || lookup_replay || lookup_valid || release_valid) begin
if (allocate_fire)
$display("%t: cache%0d:%0d mshr-allocate: addr=%0h, id=%0d, wid=%0d, PC=%0h", $time, CACHE_ID, BANK_ID,
dpi_trace("%d: cache%0d:%0d mshr-allocate: addr=%0h, id=%0d, wid=%0d, PC=%0h\n", $time, CACHE_ID, BANK_ID,
`LINE_TO_BYTE_ADDR(allocate_addr, BANK_ID), allocate_id, deq_debug_wid, deq_debug_pc);
if (fill_valid)
$display("%t: cache%0d:%0d mshr-fill: addr=%0h, id=%0d", $time, CACHE_ID, BANK_ID,
dpi_trace("%d: cache%0d:%0d mshr-fill: addr=%0h, id=%0d\n", $time, CACHE_ID, BANK_ID,
`LINE_TO_BYTE_ADDR(addr_table[fill_id], BANK_ID), fill_id);
if (dequeue_fire)
$display("%t: cache%0d:%0d mshr-dequeue: addr=%0h, id=%0d, wid=%0d, PC=%0h", $time, CACHE_ID, BANK_ID,
dpi_trace("%d: cache%0d:%0d mshr-dequeue: addr=%0h, id=%0d, wid=%0d, PC=%0h\n", $time, CACHE_ID, BANK_ID,
`LINE_TO_BYTE_ADDR(dequeue_addr, BANK_ID), dequeue_id_r, deq_debug_wid, deq_debug_pc);
if (lookup_replay)
$display("%t: cache%0d:%0d mshr-replay: addr=%0h, id=%0d", $time, CACHE_ID, BANK_ID,
dpi_trace("%d: cache%0d:%0d mshr-replay: addr=%0h, id=%0d\n", $time, CACHE_ID, BANK_ID,
`LINE_TO_BYTE_ADDR(lookup_addr, BANK_ID), lookup_id);
if (lookup_valid)
$display("%t: cache%0d:%0d mshr-lookup: addr=%0h, id=%0d, match=%b, wid=%0d, PC=%0h", $time, CACHE_ID, BANK_ID,
dpi_trace("%d: cache%0d:%0d mshr-lookup: addr=%0h, id=%0d, match=%b, wid=%0d, PC=%0h\n", $time, CACHE_ID, BANK_ID,
`LINE_TO_BYTE_ADDR(lookup_addr, BANK_ID), lookup_id, lookup_match, lkp_debug_wid, lkp_debug_pc);
if (release_valid)
$display("%t: cache%0d:%0d mshr-release id=%0d, wid=%0d, PC=%0h", $time, CACHE_ID, BANK_ID,
dpi_trace("%d: cache%0d:%0d mshr-release id=%0d, wid=%0d, PC=%0h\n", $time, CACHE_ID, BANK_ID,
release_id, rel_debug_wid, rel_debug_pc);
$write("%t: cache%0d:%0d mshr-table", $time, CACHE_ID, BANK_ID);
dpi_trace("%d: cache%0d:%0d mshr-table", $time, CACHE_ID, BANK_ID);
for (integer i = 0; i < MSHR_SIZE; ++i) begin
if (valid_table[i]) begin
$write(" ");
dpi_trace(" ");
if (ready_table[i])
$write("*");
$write("%0d=%0h", i, `LINE_TO_BYTE_ADDR(addr_table[i], BANK_ID));
dpi_trace("*");
dpi_trace("%0d=%0h", i, `LINE_TO_BYTE_ADDR(addr_table[i], BANK_ID));
end
end
$write("\n");
dpi_trace("\n");
end
end
`endif

24
hw/rtl/cache/VX_shared_mem.v vendored
View file

@ -298,22 +298,20 @@ module VX_shared_mem #(
always @(posedge clk) begin
if (!crsq_in_ready) begin
$display("%t: *** cache%0d pipeline-stall", $time, CACHE_ID);
dpi_trace("%d: *** cache%0d pipeline-stall\n", $time, CACHE_ID);
end
if (is_multi_tag_req) begin
$display("%t: *** cache%0d multi-tag request!", $time, CACHE_ID);
dpi_trace("%d: *** cache%0d multi-tag request!\n", $time, CACHE_ID);
end
if (creq_in_fire) begin
for (integer i = 0; i < NUM_BANKS; ++i) begin
if (per_bank_core_req_valid_unqual[i]) begin
if (per_bank_core_req_rw_unqual[i]) begin
$display("%t: cache%0d:%0d core-wr-req: addr=%0h, tag=%0h, byteen=%b, data=%0h, wid=%0d, PC=%0h",
$time, CACHE_ID, i, per_bank_core_req_addr_unqual[i], per_bank_core_req_tag_unqual[i], per_bank_core_req_byteen_unqual[i], per_bank_core_req_data_unqual[i],
debug_wid_st0[i], debug_pc_st0[i]);
dpi_trace("%d: cache%0d:%0d core-wr-req: addr=%0h, tag=%0h, byteen=%b, data=%0h, wid=%0d, PC=%0h\n",
$time, CACHE_ID, i, per_bank_core_req_addr_unqual[i], per_bank_core_req_tag_unqual[i], per_bank_core_req_byteen_unqual[i], per_bank_core_req_data_unqual[i], debug_wid_st0[i], debug_pc_st0[i]);
end else begin
$display("%t: cache%0d:%0d core-rd-req: addr=%0h, tag=%0h, byteen=%b, wid=%0d, PC=%0h",
$time, CACHE_ID, i, per_bank_core_req_addr_unqual[i], per_bank_core_req_tag_unqual[i], per_bank_core_req_byteen_unqual[i],
debug_wid_st0[i], debug_pc_st0[i]);
dpi_trace("%d: cache%0d:%0d core-rd-req: addr=%0h, tag=%0h, byteen=%b, wid=%0d, PC=%0h\n",
$time, CACHE_ID, i, per_bank_core_req_addr_unqual[i], per_bank_core_req_tag_unqual[i], per_bank_core_req_byteen_unqual[i], debug_wid_st0[i], debug_pc_st0[i]);
end
end
end
@ -322,13 +320,11 @@ module VX_shared_mem #(
for (integer i = 0; i < NUM_BANKS; ++i) begin
if (per_bank_core_req_valid[i]) begin
if (per_bank_core_req_rw[i]) begin
$display("%t: cache%0d:%0d core-wr-rsp: addr=%0h, tag=%0h, byteen=%b, data=%0h, wid=%0d, PC=%0h",
$time, CACHE_ID, i, per_bank_core_req_addr[i], per_bank_core_req_tag[i], per_bank_core_req_byteen[i], per_bank_core_req_data[i],
debug_wid_st1[i], debug_pc_st1[i]);
dpi_trace("%d: cache%0d:%0d core-wr-rsp: addr=%0h, tag=%0h, byteen=%b, data=%0h, wid=%0d, PC=%0h\n",
$time, CACHE_ID, i, per_bank_core_req_addr[i], per_bank_core_req_tag[i], per_bank_core_req_byteen[i], per_bank_core_req_data[i], debug_wid_st1[i], debug_pc_st1[i]);
end else begin
$display("%t: cache%0d:%0d core-rd-rsp: addr=%0h, tag=%0h, byteen=%b, data=%0h, wid=%0d, PC=%0h",
$time, CACHE_ID, i, per_bank_core_req_addr[i], per_bank_core_req_tag[i], per_bank_core_req_byteen[i], per_bank_core_rsp_data[i],
debug_wid_st1[i], debug_pc_st1[i]);
dpi_trace("%d: cache%0d:%0d core-rd-rsp: addr=%0h, tag=%0h, byteen=%b, data=%0h, wid=%0d, PC=%0h\n",
$time, CACHE_ID, i, per_bank_core_req_addr[i], per_bank_core_req_tag[i], per_bank_core_req_byteen[i], per_bank_core_rsp_data[i], debug_wid_st1[i], debug_pc_st1[i]);
end
end
end

8
hw/rtl/cache/VX_tag_access.v vendored
View file

@ -68,16 +68,16 @@ module VX_tag_access #(
always @(posedge clk) begin
if (fill && ~stall) begin
if (is_flush) begin
$display("%t: cache%0d:%0d tag-flush: addr=%0h, blk_addr=%0d", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(addr, BANK_ID), line_addr);
dpi_trace("%d: cache%0d:%0d tag-flush: addr=%0h, blk_addr=%0d\n", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(addr, BANK_ID), line_addr);
end else begin
$display("%t: cache%0d:%0d tag-fill: addr=%0h, blk_addr=%0d, tag_id=%0h, old_tag_id=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(addr, BANK_ID), line_addr, line_tag, read_tag);
dpi_trace("%d: cache%0d:%0d tag-fill: addr=%0h, blk_addr=%0d, tag_id=%0h, old_tag_id=%0h\n", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(addr, BANK_ID), line_addr, line_tag, read_tag);
end
end
if (lookup && ~stall) begin
if (tag_match) begin
$display("%t: cache%0d:%0d tag-hit: addr=%0h, wid=%0d, PC=%0h, blk_addr=%0d, tag_id=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(addr, BANK_ID), debug_wid, debug_pc, line_addr, line_tag);
dpi_trace("%d: cache%0d:%0d tag-hit: addr=%0h, wid=%0d, PC=%0h, blk_addr=%0d, tag_id=%0h\n", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(addr, BANK_ID), debug_wid, debug_pc, line_addr, line_tag);
end else begin
$display("%t: cache%0d:%0d tag-miss: addr=%0h, wid=%0d, PC=%0h, blk_addr=%0d, tag_id=%0h, old_tag_id=%0h", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(addr, BANK_ID), debug_wid, debug_pc, line_addr, line_tag, read_tag);
dpi_trace("%d: cache%0d:%0d tag-miss: addr=%0h, wid=%0d, PC=%0h, blk_addr=%0d, tag_id=%0h, old_tag_id=%0h\n", $time, CACHE_ID, BANK_ID, `LINE_TO_BYTE_ADDR(addr, BANK_ID), debug_wid, debug_pc, line_addr, line_tag, read_tag);
end
end
end

16
hw/simulate/simulator.cpp
View file

@ -5,6 +5,10 @@
#define ENABLE_MEM_STALLS
#ifndef TRACE_DELAY
#define TRACE_DELAY 0
#endif
#ifndef MEM_LATENCY
#define MEM_LATENCY 24
#endif
@ -24,7 +28,9 @@
#define VL_WDATA_GETW(lwp, i, n, w) \
VL_SEL_IWII(0, n * w, 0, 0, lwp, i * w, w)
uint64_t timestamp = 0;
uint64_t sim_trace_delay = TRACE_DELAY;
static uint64_t timestamp = 0;
double sc_time_stamp() {
return timestamp;
@ -120,7 +126,9 @@ void Simulator::step() {
void Simulator::eval() {
vortex_->eval();
#ifdef VCD_OUTPUT
trace_->dump(timestamp);
if (timestamp >= sim_trace_delay) {
trace_->dump(timestamp);
}
#endif
++timestamp;
}
@ -141,11 +149,11 @@ void Simulator::eval_mem_bus() {
bool has_response = false;
// schedule memory responses in FIFO order
// schedule memory responses that are ready
for (int i = 0; i < MEMORY_BANKS; ++i) {
uint32_t b = (i + last_mem_rsp_bank_ + 1) % MEMORY_BANKS;
if (!mem_rsp_vec_[b].empty()
&& (0 == mem_rsp_vec_[b].begin()->cycles_left)) {
&& (mem_rsp_vec_[b].begin()->cycles_left) <= 0) {
has_response = true;
last_mem_rsp_bank_ = b;
break;